Speech Recognition Acoustic Model Frame Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems face challenges in efficiently calculating pronunciation probabilities due to the increased time required for processing longer speech units, especially when using Bidirectional Recurrent Deep Neural Networks (BRDNNs), which consider extra frame information and temporal considerations.
Innovation Solution
A speech recognition apparatus and method that extracts select frames from a speech signal using a preprocessor and calculates acoustic scores using a DNN-based acoustic model, allowing for the generation of a second speech signal with reduced frames, which are then used to infer scores for the original speech, thereby reducing computational time without performing full acoustic modeling on all frames.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Bidirectional Recurrent Deep Neural Network (BRDNN) is used to calculate pronunciation probabilities considering all frames, then accuracy is improved, but processing time increases significantly
Solution Approach 1:
The patent divides the speech signal into multiple segments or frames, and selectively processes only certain frames through the BRDNN acoustic model rather than all frames. This segmentation approach maintains accuracy for critical frames while reducing overall computational time by excluding less important frames from full acoustic modeling.
Solution Approach 2:
The patent applies partial action by performing complete acoustic modeling only on selected frames that are most critical for recognition accuracy, while using simplified or interpolated methods for remaining frames. This selective processing maintains sufficient accuracy while significantly reducing the total computational burden.
2Reliability
If acoustic modeling is performed on all frames of speech, then recognition accuracy is maintained, but computational complexity increases
Solution Approach 1:
The speech signal is segmented into multiple frames, and the patent applies different processing levels to different segments. Critical frames undergo full acoustic modeling while less critical frames use simplified processing, thereby reducing computational complexity while maintaining overall recognition reliability.
Solution Approach 2:
Different quality levels of acoustic processing are applied to different portions of the speech signal based on their importance. Frames containing critical phonetic information receive full BRDNN processing, while other frames use reduced processing, optimizing the balance between reliability and computational complexity.
Data Source
AI summary
Provided is a speech recognition apparatus. The apparatus includes a preprocessor configured to extract select frames from all frames of a first speech of a user, and a score calculator configured to calculate an acoustic score of a second speech, made up of the extracted select frames, by using a Deep Neural Network (DNN)-based acoustic model, and to calculate an acoustic score of frames, of the first speech, other than the select frames based on the calculated acoustic score of the second speech.